1. Reminder: Please user our affiliate links to get to your favorite stores for holiday shopping!

Noise on Hi Res shots from E-M5ii and Pen F

Discussion in 'Olympus Cameras' started by pdk42, Apr 17, 2016.

  1. pdk42

    pdk42 One of the "Eh?" team

    Jan 11, 2013
    Leamington Spa, UK
    In anticipating the E-M1 Mk ii release later this year, I've been musing about whether the hand-holdable Hi Res rumours will prove to be true or not. If they are true, then it should deliver a significant step up in IQ for many types of photography (landscapes, cityscapes etc). I know it'll deliver higher res, and apparently more accurate colour - but what about noise? I can see that just by having lower magnification to the final image that noise will improve, but I suppose at the pixel level nothing changes.

    So, I'd be interested to hear from any E-M5ii or Pen F users what their experiences are with noise on Hi Res shots. Have you tried it at higher ISOs?
     
  2. Klorenzo

    Klorenzo Mu-43 All-Pro

    Mar 10, 2014
    Lorenzo
    I would expect better noise both in the hi-res version and in the downscaled version. You have much more samples for each location (8 shots to cover 4 times the "pixels") so all the sources of random noise should mostly cancel each other out. When you downscale to a significantly smaller size, like 16/24MP, this average effect is even stronger. But I've never seen a real comparison about this. The hi-res is also limited to ISO 1600.
     
  3. pdk42

    pdk42 One of the "Eh?" team

    Jan 11, 2013
    Leamington Spa, UK
    Yes, that confirms pretty much what I'm thinking too. It would be good to see some samples though and/or thoughts fdom those who have used it.
     
  4. Machi

    Machi Mu-43 Veteran

    206
    May 23, 2015
    Here is quick comparison between different modes of E-M5II and Pen-F at ISO1600.
    Pixel shift images were binned by factor 2 to same size as standard images.
    All images are from the DPreview image comparison tool.

    Comparison.
     
    • Informative Informative x 2
    • Useful Useful x 2
  5. wjiang

    wjiang Mu-43 Hall of Famer

    Wow, pixel shift is very smooth.
     
    • Agree Agree x 1
  6. MichaelSewell

    MichaelSewell Mu-43 Top Veteran

    507
    Sep 1, 2015
    Burnley, UK
    Michael A. Sewell
    :thiagree:
     
  7. Jonathan F/2

    Jonathan F/2 Mu-43 Hall of Famer

    Aug 10, 2011
    Los Angeles, CA
    I actually find the images a bit noisy even at base ISO when using hi res mode. If you're shooting jpeg, I recommend using at least 'normal' noise reduction or if in raw have a custom noise reduction preset. It's most noticeable in night exposures. I've compared shots with my old FF D600 and E-M5 II in high res mode during evening conditions and the 24mp FF sensor is a bit cleaner. In good light though, it's not as noticeable and easily mendable in raw processing. My only gripe with high res mode though, is that it's capped at 8 seconds and at f/8. You need at least f/11 to get the really nice light points when shooting at night.

    The one benefit of the hi res mode is bringing a small rig that's great for tripod land/urban scapes and then being able to use the camera in normal shooting mode for street and portraiture at the regular 16/20mp. Works great for travel.
     
  8. MNm43

    MNm43 Mu-43 Regular

    103
    Mar 19, 2014
    I played with a Pen-F at a store for 20 or so minutes so take this for what it is worth. I fixed the ISO at 200 and did one shot with with pixel shift and one without. Looking on the back display, my impressions matched the images presented above. My concern with this technology is that in most landscape scenes, the fine details you are trying to capture (leaves, branches, flags distant and otherwise) are exactly those details likely to move in the slightest breeze and end up smeared. Even with close scenes, grass and flower petals quiver in the least breeze. If you want to get a sense of how problematic this is, try using a program like PhotoAcute where you can use frames fired off at 10+ fps. It doesn't take much to find problems in the files.

    I do a lot of panoramas, and one thing I have to say is that the processing software has gotten very sophisticated at blending overlapped images as well as doubled images. My sense is that pixel shift is a kind of 'dumb' processing that simply combines the pixels without doing any image analysis. Whether the kind of image analysis processing that panoramic software uses can be applied to pixel shift images (or even whether you'd want to) might be an interesting question.

    Now if we are talking studio work, then I agree it is pretty amazing technology.
     
  9. Steven Norquist

    Steven Norquist Mu-43 Regular

    45
    Jan 24, 2016
    Steven Norquist
    Here are two shots taken with the Pen-F.
    Both were shot raw.
    The first, normal 20mp, the second, 80mp then sized down to 20mp.
    100% crops.
    Look at two things, resolution and noise.
    What do you see?

    01_Standard 20mp_2. 02_80mp sized to 20mp_2.
     
  10. MNm43

    MNm43 Mu-43 Regular

    103
    Mar 19, 2014
    Better detail on the fence, the shingles, and (plaster?) wall. The noise looks better controlled in the hi-res as well. The branches in the 80mp are problematic. I don't see much difference in detail in the distance - whether that is because it is out of the DOF or atmospheric 'noise' is hard to say.
     
  11. MaK543

    MaK543 Mu-43 Regular

    139
    May 1, 2012
    MD USA
    The hi-res shot easily wins in term of resolution and noise in this comparison. Look the window! The tree branches have motion blur which is expected for moving objects in 'hi-res' mode.
     
    • Agree Agree x 1
  12. Ctein

    Ctein Mu-43 Regular

    Dear Folks,

    Machi's postings accord with my noise results (Jonathan, you're situation is anomalous in some ways-- I'll get back to it). The HR files are essentially noise-free. Not that the E-M5 is all that noisy at ISO 200, normally, it's already very good. In HR mode, even that miniscule amount of noise is entirely gone; the tonality is absolutely smooth with no hint of variation from pixel to pixel. It's amazing. It extends all the way down to the shadows, which never get grainy. If you feel the need to open up the shadows, they stay nice and creamy. It is very, very much a large-format look, almost a large-format contact-print look.

    At the highest ISO allowed for HR, ISO 1600, the HR noise level is about the same as what you'd see in a normal exposure at ISO 200.

    MNm43, you're wrong about what the pixel shift software is doing. The maths of subsample processing are precise and well-understood. The software is performing a differential analysis on the eight frames, which are offset in one half-pixel increments. By doing that, the software can not only determine the luminance values at half-pixel resolution, but it collects full three-color information for each and every one of those half-pixels. That is one of the big reasons why the resultant noise drops so much. A lot (not all) of the low-ISO noise you see in processed raw files is a side effect of the interpolation algorithms that generate full-color images from adjacent single-color pixels. That interpolation doesn't have to be done with an HR image set.

    It also has the side effect of improving color rendition with some seriously-weird light sources, like blue and violet LED Christmas lights, which often give the RAW conversion/interpolation algorithms fits.

    The thing is, for this to work well, the positioning of those samples has to be really precise, way better than one half pixel. That's why any subject movement or the least bit of camera shake generates the weird crosshatch patterns in the processed file. Think of it as a kind of aliasing that occurs when the movement error is on the same scale as the sampling distance.

    I am hopeful that the 1/60th sec HR mode in the improved E-M1 will be sufficiently short to allow handholding with a decent array of lenses, but it's going to have to be subject to test. I mean, I can handhold my 45mm lens down to a 20th of a second and get pixel perfect sharpness. But for HR to work well, I'm going to need more like 1/8th-pixel** perfect sharpness. Can I do that at 1/60th sec? I dunno!

    Differential analyses like REALLY clean data to start with. They are sensitive to deviations from the ideal. That's the reason why HR mode is limited to ISO 1600. Push beyond that and the amplification noise from the camera starts to screw with the differential analysis.

    This relates to what you observe, Jonathan, even though you were talking about working at ISO 200. At an 8 second exposure, you're well up into the range where long-exposure sources of noise (E.g., thermal electrons) start to become an issue. The auto-noise-reduction software in the camera doesn't deal with that kind of noise, because mostly it truly is random noise–– it'll vary from frame to frame. You can't subtract most of it out. Just like high ISO amplification noise, this will propagate through the differential algorithms to some degree.

    A 3 second exposure at ISO 200 does look slightly noisier than a normal length exposure, but you really would have to be a fetishist to care. It is invisible in even a 24 x 32" print.

    ~~~~

    Off on a tangent, both normal and HR RAW files benefit from a bit deconvolution sharpening, but it works much better on the HR photos. There's a distict improvement in sharpness in the HR file and edge acutance jumps considerably without getting crunchy or, well, “edgy.” It just looks naturally sharper. It's cruftier on a normal photograph. The HR edge detail just rolls off smoothly, like large-format film with a decent lens.

    ~~~~

    (** I'm pulling that number out of my, um,… hat. I'm not doing the maths. My point though is that the positioning has to be LOTS more stable than merely one half-pixel.)

    pax \ Ctein
    [ Please excuse any word-salad. MacSpeech in training! ]
    ======================================
    -- Ctein's Online Gallery Ctein's Online Gallery
    -- Digital Restorations Photos Restored Digitally by Ctein
    ======================================
     
    Last edited: Apr 23, 2016
    • Like Like x 5
    • Informative Informative x 5
  13. Steven Norquist

    Steven Norquist Mu-43 Regular

    45
    Jan 24, 2016
    Steven Norquist
    Quick note on the comparison shots I posted: Focus was set manually on the distant antenna on the mountain.
     
  14. Jonathan F/2

    Jonathan F/2 Mu-43 Hall of Famer

    Aug 10, 2011
    Los Angeles, CA
    In order to get somewhat interesting urban night shots especially with cars driving and light streaks, those long exposures are necessary. Also in order to get nice light points, going smaller than f/8 is necessary. I guess I'm pushing high res mode in an area it wasn't really suited, but the files are great when done right. The random noise isn't such an issue now after customizing my own NR preset in Adobe Camera Raw and it's great that ACR can read the .ORF files for high res mode.

    I think where high res mode can really do it's magic is at sunrise or dusk where the exposures don't need to be as long. Some of my most beautiful high res mode shots have been shot during that time.
     
  15. Ctein

    Ctein Mu-43 Regular

    Dear Jonathan,

    Oh, I wasn't faulting your intentions. You know the kind of photos you want to make. It's as you said, though-- you're trying to push the camera beyond some of the reasonable limits.

    I'm curious, though-- what lens(es) are you using that you need to stop down beyond f/8 to get good pinpoints? A "starfield" test is the absolutely toughest one you can throw at a lens, no question. But, most lenses, if they haven't cleaned up by f/8, they ain't gonna at f/11.

    FWIW, the f/8 limit is an arbitrary one in the camera software-- it's not driven by the differential noise issues. I'd *guess* that Olympus did that because too many photographers don't understand diffraction limits and optimum apertures, and they wanted to keep them out of trouble. It's not the design decision I'd have made, but such is life. Maybe they'll change it in the future.

    It's also a little weird that it defaults to JPEG only and the JPEG is smaller than the RAW file, because the Olympus engineers don't think there's more detail to be gotten from it. That's sure not my experience. Unsharpened, maybe so, but judicious deconvolution works wonders, if sheer resolution is what you're after. Me, I'm finding resolution is one of the least interesting reasons for using the HR mode; whodathunk?!

    I note that the RAW+JPEG mode does save a single normal-res frame as well, as insurance in case the HR fails. It's got an ORI designation, which ACR and DNG Converter can't understand, but other RAW processors, like Photo Ninja and Iridient, can. It's nice to have the backup.

    pax / Ctein
     
    Last edited: Apr 24, 2016
  16. Klorenzo

    Klorenzo Mu-43 All-Pro

    Mar 10, 2014
    Lorenzo
    I try to elaborate a little on this. Each "pixel" of the sensor is single-color (due to a micro colored filter applied on top), instead digital images pixels have (in general) three colors, like those used in a computer monitor. So to create a 16MP image with three full colors for each pixel out of 16MP single-color ones you just have to made up 2/3(!) of the final information. You do this with a smart copy-and-past from the surrounding pixels i.e. the demosaicing algorithms (there are a few of these: AMaZE, PPG, etc.).
    According to this article, in hi-res mode the sensor first does a full pixel "scan" where it collects a red, blue and two greens readings for each pixel. There is no "real" shift here: the sensor moves just to place a pixel with a different "color" in the same location of the projected image.
    Then it does a second "half-pixel" scan to create the extra resolution again with full color values for the virtual pixels (here you are actually seeing different parts of the scene). I suppose some kind of "smart smoothing" is done with these eight samples too.

    There is some overlap in the pixels and this is why they prefer to create a 40MP JPEG rather then a 64MP one. I do not know why they did not go for a full-pixel shift in the second scan, maybe they preferred to favor accuracy over extreme resolution.

    In theory they could also make a 16MP four shots only mode (full RGB), a 16MP eight-shots (full RGB with two readings), etc.